When I first started building a product using LLM APIs, I adopted LangGraph. Later, I moved away from it and implemented chat history and state management myself. Here’s what I learned through that process.

LangGraph — a Lifesaver at First

When I first needed to build a chatbot, LangGraph was a huge help. The goal was to create a chatbot that could respond based on existing data — something like a banking app chatbot. At that time, I had no experience with similar systems. The chatbot needed to extract appropriate search keywords from user questions, query the database, and then format the results as proper answers — all of which were difficult to design.

Building a structured and predictable product using an LLM API is challenging because, unlike traditional APIs, LLM outputs are non-deterministic. In this post, I’ll share what I learned while developing a chatbot that manages financial assets and how I made it stable.

Let’s assume a sample chatbot that manages two assets: USD and KRW. The user can check balances, transfer funds, or exchange currencies.

Building a Prototype with LangGraph

Using LangGraph, it’s easy to build a working chatbot prototype. With create_react_agent, you can create an agent by simply adding a prompt and a few tools. For example, the prompt could be:

Coroutine allows you to implement the primary dialogue system intuitively.

Dialogs are used in so many games. Many games use dialogs. Visual novel, as well as RPG, action games, puzzle games, and all other games can have characters that speak, and they use the dialogue to provide information to the player. Also, users are familiar with the dialogue system.

However, Unity does not provide a dialogue system. A programmer should buy an asset or manually implement it. If the requirements for the dialog are complicated, you will use a feature-rich and complex paid asset, but if you have a simple need, it is better to create it yourself.